Lin et al. EURASIP Journal onWireless Communications andNetworking (2016) 2016:71 DOI 10.1186/s13638-016-0572-x

RESEARCH Open Access

A separate design principle forpriority-aware packet collection in industrialcyber-physical systemsFeilong Lin1,2, Cailian Chen1,2*, Qimin Xu1,2, Cunqing Hua3 and Xinping Guan1,2

Abstract

Industrial cyber-physical systems (ICPS) bridge the physical factory floor and the cyber computational space byleveraging the emerging techniques, such as wireless sensor networks for ubiquitous connection and perception,where sensors are deployed to monitor and collect industrial data according to different timeliness requirements.Priority-aware schemes are often used to tackle the differentiated data packet collection. However, it is challenging tocoordinate the multichannel access while meeting the priority-aware transmission requirement in ICPS. In this paper,we propose a separate design principle (SDP) for priority-aware packet collection. In SDP, the transmission of eachpriority class of sensors is separately scheduled by a multichannel superframe which fully uses the available channels.As a result, each time slot on each channel is repeatedly scheduled to sensors of different priorities by differentsuperframes. Then, a priority-aware transmission mechanism is devised to coordinate the transmissions of differentsensors in the predefined priority order. Simulation results show that for four priorities, SDP achieves as low as 16, 18,23, and 36 % of the mean packet transmission delay of non-overlap TDMA scheduling for each priority class,respectively. Moreover, SDP greatly outperforms IEEE 802.15.4e protocol for packet collection with two priorities. Wealso demonstrate the feasibility of SDP based on the implementation of software-defined radios.

Keywords: Industrial cyber-physical systems, Priority-aware packet collection, Multichannel scheduling, Separationdesign principle

1 IntroductionCyber-physical systems (CPS) are an integrated infras-tructure that involves sensing, computation, commu-nications, and control [1, 2]. Manufacturing industryintegrated with CSP, aka industrial cyber-physical sys-tems (ICPS), bridges the physical factory floor and thecyber computational space by leveraging the emergingtechniques, such as wireless sensor networks (WSNs) forubiquitous connection and perception [3]. ICPS have beenconsidered as the landmark in the development of Indus-try 4.0, the next generation of manufacturing industry. Itis reported by GE that about 46 % of the global economy

*Correspondence: cailianchen@sjtu.edu.cn1Department of Automation, Shanghai Jiao Tong University, 800, DongchuanRd., Minhang District, 200240 Shanghai, People’s Republic of China2Key Laboratory of Systems Control and Information Processing, Ministry ofEducation of China, 800, Dongchuan Rd., Minhang District, 200240 Shanghai,People’s Republic of ChinaFull list of author information is available at the end of the article

or $32.3 trillion in global output can benefit fromICPS [4].In ICPS scenario, sensors are deployed and various

monitoring data are to be collected according to the dif-ferent timeliness requirements. For example, industrialprocess monitoring data transmission for online controlcan only tolerate the latency nomore than tens of millisec-onds [5]. Thus, they require higher transmission prioritythan other monitoring data which may allow long latency,such as the machine health monitoring data. Priority-aware schemes are often used to tackle the differentiateddata transmission. However, it is still challenging to coor-dinate the multichannel access while meeting the priority-aware transmission requirement due to the complexity ofcollision avoidance and transmission prioritization.Since the IEEE Std 802.15.4 [6] was released, which

defines the MAC and PHY specifications in ISM (indus-trial, scientific, and medical) band, several standardized

© 2016 Lin et al. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 InternationalLicense (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in anymedium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commonslicense, and indicate if changes were made.

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 2 of 14

protocols have been successively released, such as Zigbee[7], WirelessHART [8], ISA100.11a [9], and WIA-PA[10]. Zigbee, utilizing CSMA/CA (carrier sense multi-ple access/collision avoidance) mechanism, provides thecapability of self-organization and scalability. However,it is not efficient enough for large-scale networks. Asshown in [1, 11], when more than 20 nodes in the net-work, the packet delivery ratio falls below 40 % andthe packet transmission delay exceeds 100 ms. Hence,CSMA/CA is not suitable to serve the ICPS with a largenumber of devices and strict timeliness requirement. Asan alternative solution, TDMA (time-division multipleaccess), which achieves the deterministic and predictabletransmission, is accepted as the main scheduling mech-anism by WirelessHART, ISA100.11a, and WIA-PA. Bythis reservation-based scheduling approach, these threeprotocols have specified four transmission priorities fordifferent data, i.e., command, process data, normal, andalarm from the highest priority to the lowest. However,TDMA also faces two folds of deficiencies. Firstly, it can-not provide timely access for the transmission requests,especially in the random event-driven scenario. Second,since each slot is exclusively scheduled to one sensor, itmay lead to waste if the sensor does not have data totransmit.In this paper, we propose a flexible and efficient TDMA

scheduling mechanism for priority-aware packet collec-tion in ICPS. Consider the industrial WSN consistingof multiple classes of sensors, where each class of sen-sors belongs to different transmission priorities. The datapackets from higher priority sensors have transmissionsuperiority over the packets from lower priority sensors.We introduce a separate design principle (SDP) based onthe multichannel superframe used in IEEE 802.15.4e [12].The main idea of SDP is to separately design multichan-nel superframe for each class of sensors by fully utilizingthe channels in the network. As a result, each time sloton each channel is repeatedly scheduled to sensors ofdifferent priorities by different superframes. Then, a pri-ority coordinator is devised to manage the transmissionaccording to the priority order. At the beginning of eachslot, certain sub-slots are reserved for sensors to indi-cate their transmission priority. By checking the priorityindicators, the lower priority sensors can opportunisti-cally utilize the unused slots which have been scheduledto higher priority sensors. We preliminarily reported theSDP with two priority classes in [13]. In this paper, weextend the SDP to multipriority packet collection scenarioand present comprehensive theoretical analysis for SDP.The main contributions of this work are summarized asfollows:

• A separate design principle is proposed formultichannel scheduling in priority-aware packet

collection. SDP not only guarantees the transmissionpriority of different sensors but also decreases thetransmission delay of the lower priority sensors byallowing them to reuse the slots scheduled to higherpriority sensors opportunistically.

• Noting that the periodic transmission benefits thedecrease of packet transmission delay, a greedymultichannel superframe determination (GMSD)algorithm is devised to optimize the multichannelsuperframe design.

• The lower bound of mean packet transmission delayfor each class of sensors is derived. It is also provedthat the SDP-based scheduling obtains lower meanwaiting delay than non-overlap TDMA scheduling.

The remainder of this paper is organized as fol-lows: Section 2 presents a simple review of the relatedworks. Priority-awaremultichannel scheduling problem isdescribed in Section 3. The main results of this work arederived in Section 4. Section 5 shows the performanceanalysis for the proposed SDP. Finally, simulations andexperiments are conducted in Section 6 and the paper isconcluded in Section 7.

2 Related worksPriority-aware packet collection is a fundamental tech-nique to coordinate the sensors’ data packet transmis-sion with different timeliness requirements in ICPS. IEEEStandard 802.15.4, extensively applied in industrial appli-cations, has provided the preliminary solution for packettransmission with different delay constraints [6]. Byreserving the guaranteed time slots (GTSs) in contention-free period (CFP), the high-priority data packets can betransmitted in a collision-free way. In contrast, the low-priority data packets are committed to access to chan-nel by CSMA-CA approach in contention access period(CAP). However, for large-scale network, the transmis-sion efficiency is very low [11]. Based on IEEE 802.15.4protocol, an adaptive strategy is proposed in [14] to makethe tradeoff between CFP and CAP, thus to deal with thedynamics of high- and low-priority traffics. Comparedto reserved transmission for high-priority packets, theauthors in [15] propose to transmit high-priority pack-ets in event-driven manner while the low-priority packetsin TDMA manner. By allowing high-priority packet tohijack the transmission chance of low-priority packet,this approach guarantees the transmission superiority ofhigh-priority packets. It works well in the scenario wherethe high-priority packet rate is very low, and no colli-sion would happen among the transmissions of the high-priority packets. The authors in [16, 17] do not considerthe absolute packet transmission priority; instead, theypropose a proportional delay model and a transmissionscheduler based on queueing information. As a result,

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 3 of 14

the mean transmission delays of different data trafficsapproach a proportional fashion. In [18–20], the packettransmission priority from the perspective of remainderhops to the destination node is considered, and then, thedynamic scheduling methods are proposed to minimizethe end-to-end delay.The studies in above literatures are specified on single

channel, which necessitate the multichannel access coor-dination to extend to multichannel industrial WSNs. Toexploit the multichannel diversity, the authors in [21] pro-pose a multichannel superframe scheduling mechanismfor cluster-tree topology network based on IEEE 802.15.4protocol. It allocates each cluster with one orthogonalchannel to avoid inter-cluster interference. The super-frame is still defined on single channel. A fixed prioritypacket transmission method based on WirelessHART isproposed in [22] for multichannel multihop networks. Tomeet the delay constraint, each sensor determines thetransmission order of packets with the consideration oftheir remaining hops to the destination. Multiple channelsare used to support parallel transmissions over multiplerouting paths to reduce the end-to-end delay of packetdelivery. In [23], the authors considered the multiprior-ity multichannel access in cognitive radio environment.Similar to [15], high-priority traffics are allowed to pre-empt transmission chances of low-priority traffics. Theauthors in [24] proposed a distributed priority-awaremul-tichannel access scheme based on weighted congestiongame, where higher priority traffics are assigned withlarger weight coefficients, thus to acquire more spectrumresource.For priority-aware packet collection in industrial field,

the access point (AP) cannot easily acquire the timelyqueueing states of sensors without efficient channel accessmechanism. Hence, the TDMA scheduling, a reservation-based mechanism, is an appropriate choice [5]. We aimto reduce the packet transmission delay of each prior-ity class of sensors by designing separate multichannelsuperframes and priority-aware transmission coordina-tion mechanism. Further, it is worth to note that for eachsensor, the evenness of the distribution of scheduled slotstakes effect on the packet collection delay. To the best ofour knowledge, this issue has not been considered in theexisting works. In this paper, we explore the effect of theevenness on transmission delay and develop an algorithmto optimize the multichannel superframe design.

3 Problem description3.1 Network modelsConsider a star topology network in the industrial field.The sensors are deployed to monitor the industrial pro-cess and one access point (AP) is appointed to collect thedata packets from the sensors. Due to different timeli-ness requirements, the sensors are classified to C classes

with different transmission priorities. Suppose that eachclass c consists of Nc sensors and the sensors in class care assigned with higher transmission priority over anyclass c′ > c. Similar to some existing studies [15, 23],the data packet flows generated by each class of sen-sors follow independent Poisson processes with packetgenerating rate {λn,c|n = 1, . . . ,Nc, c = 1, . . . ,C}. Sup-pose that the rates are slowly time-varying. For example,they keep invariable during the production of one batchor one order. The packet size is fixed and normalizedto 1.Suppose that the network operates on L channels. Time

is slotted and the network is synchronized. During eachtime slot, one packet can be transmitted on each channel.The packet transmission of each sensor follows first-comefirst-served manner. Each sensor only transmits on a sin-gle channel in a slot, but can do per-slot channel hopping ifrequired. AP supports multichannel transmission, whichcovers all L channels.To guarantee the transmission schedulability of the net-

work, the average packet arrival rate in the network shouldbe less than L. Accordingly, the network traffic rate mustsatisfy

∑Cc=1

∑Ncn=1 λn,c < L.

3.2 Priority-aware multichannel scheduling problemThe purpose of this paper is not only to coordinate themultipriority packet collection in multichannel networksbut also to reduce the packet transmission delay for eachpriority class of sensors. To this end, the following twosub-problems are considered.

3.2.1 Optimalmultichannel superframe designIt is observed that the TDMA-based multichannelscheduling would affect the packet transmission delay.The evenness of the distribution of scheduled time slotsfor each sensor, in particular the second moment ofscheduling intervals, affects themean packet transmissiondelay, which will be proved in Section 4.1.2. Therefore, thedelay optimization should be considered in themultichan-nel superframe design.

3.2.2 Priority-aware transmission coordinationIn order to meet the differentiated packet collectionrequirements, a priority-aware transmission coordina-tion mechanism is needed. This coordination mechanismmanages the transmission of different sensors in the pre-defined priority order. Besides the resolution of transmis-sion priority, the transmission delay of each priority classof sensors can be further improved by devising propertransmission coordination mechanism.

4 Main resultsThis section first presents the optimal multichannelsuperframe design. Based on multichannel superframe

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 4 of 14

design, the SDP for priority-aware packet collection isthen derived.

4.1 Multichannel superframe designRecently, IEEE 802.15.4e [12] has been released, in whichthe multichannel superframe is defined by the determin-istic and synchronous multichannel extension (DSME)mechanism. To solve the first problem presented inSection 3.2.1, we use TDMA-basedmultichannel schedul-ing for the superframe design, such as DSME-GTS allo-cation in CFP in IEEE 802.15.4e protocol. The CAPcan be used to broadcast beacon frame to deliver thescheduling information to the sensors, and which canbe set to a small period, e.g., one time slot. This workmainly considers the transmission scheduling in CFPand will not refer to the phase of the transmission inCAP.

4.1.1 TDMA-basedmultichannel superframe designThe basic idea of SDP is to separately design a multichan-nel superframe for each priority class of sensors. Withoutloss of generality, consider the multichannel supframedesign for a general priority class with N sensors (the pri-ority order c is omitted in this section). The packet arrivalsof sensors are Possion-distributed with rates {λn|n =1, . . . ,N}. Denote A as the multichannel superframe onL channels over a period of T slots, which can be repre-sented by a L×T matrix. Each element ai,j is one resourceblock of the jth slot on the ith channel. All the resourceblocks of A are allocated to sensors in a proportional way.By taking the packet rate of each sensor as the propor-tional coefficient, the portion of resource allocated to eachsensor is

Rn = λnU

LT , (1)

where U = ∑Nn=1 λn is the overall packet rates of the sen-

sors. In implementation, Rn is set to be an integer number,and

∑Nn=1 Rn = LT .

In this context, the multichannel superframe designproblem is to allocate the LT resource blocks toN sensorsproperly. For simplicity, we assign an integer to each ai,j torepresent the resource block allocation in A, e.g., ai,j = nmeans that jth slot on ith channel is allocated to sensorn. In practice, the whole superframe information is notnecessary for each sensor. To save the storage resource,each sensor only needs to store its transmission schedule,denoted by An = {(i, j)|ai,j = n}. Note that An is a two-dimensional array. Obviously, the length of the array An isRn. Since no more than one channel can be allocated to asingle channel sensor at each time slot, An has a strictlyincreasing order with respect to time slot index j.

4.1.2 Mean packet transmission delayDenote {tn,k|k = 1, · · · ,Rn} as the time slot indexes intransmission scheduleAn. Then, with regard to the super-frame A, the scheduling interval for sensor n is defined as

sn,k ={tn,k+1 − tn,k , k = 1, . . . ,Rn − 1,T − tn,k+1 + tn,k , k = Rn.

(2)

When the sensor delivers data packets according to thesuperframe periodically, the first and second moments of{sn,k} are

sn = TRn

, (3)

s(2)n = 1Rn

Rn∑k=1

(sn,k)2. (4)

The packet transmission of each sensor can be mod-eled as a queueing process with Poisson arrival. Then,the packet transmission follows as the general distributionwith the mean transmission interval sn and the secondmoment s(2)n . According to M/G/1 queueing model [25],the mean waiting time of sensor n can be obtained asfollows according to Pollaczek-Khintchine formula:

Wn = λns(2)n2(1 − λnsn)

. (5)

To emphasize the influence of scheduling interval on themean waiting delay, we have the following lemma.

Lemma 1. For Poisson-distributed packet arrival, alower second moment of scheduling interval achieves alower mean waiting delay. The lowest mean waiting delayis achieved by periodic scheduling, i.e.,

WLBn = λnT2

2Rn2 − 2λnTRn, (6)

and the lower bound of mean packet transmission delay(including the waiting slots and the transmission slot) ofsensor n is

DLBn = WLB

n + 1. (7)

Proof. From (5), a smaller s(2)n results in smaller Wn.Note that the minimum s(2)n is achieved when all resourceblocks assigned to the sensor n are equally distributedin time-dimension, i.e., s(2)n =

(TRn

)2. Hence, periodic

scheduling achieves the lowest mean waiting delay asshown by (6). Since one packet is delivered only in onetime slot, after waiting, the packet is transmitted withinone slot. Therefore, (7) holds for the lower bound of meanpacket transmission delay.

As the secondmoment cannot be easily determined, theformulation of lower bound will be used in the following

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 5 of 14

analysis. For simplicity, the superscript “LB” is removed,e.g., using Dn instead of DLB

n .

4.1.3 Determination ofmultichannel superframeIn the following, we present the determination of multi-channel scheduling A. With Rn resource blocks allocatedto sensor n, the mean scheduling interval sn is fixedaccording to (3). From (5), it is shown that a smaller sec-ond moment of scheduling interval s(2)n will provide alower mean waiting delay as well as mean transmissiondelay in (7). Hence, a good superframe should make sec-ond moment of scheduling intervals as small as possible.In this paper, with the consideration of fairness amongsensors, we try to find the optimal schedule A to satisfy

A∗ = argminmaxn∈[1, N]

s(2)n . (8)

Problem (8) is an integer programming problem, whichnormally does not have an analytical solution. To this end,we design a greedy multichannel superframe determina-tion (GMSD) algorithm as shown in Algorithm 1.

Algorithm 1Greedy multichannel superframe determination1: Input: L; T ; {λn|n = 1, . . . ,N};2: Output: A;3: Calculate {Rn|n = 1, . . . ,N} according to (1);4: // Initial scheduling:5: B1×LT ← 0; K ← 0;6: for n = 1 to N do7: B(K + 1 : K + Rn) ← n;8: K ← K + Rn;9: end for

10: Reshape: AL×T ← B1×LT ;11: // Optimal scheduling:12: flag ← 1;13: while flag = 1 do14: s(2)max ← max

{s(2)n

}; n ← argmax

{s(2)n

}; Jn,k ←

argmax{sn,k};15: Select ai,j with j = Jn,k , ai,j = n;16: Temp1×L ← 0;17: for l = 1 to L do18: Exchanging ai,j and al,j+1; // Try19: Temp(l) ← s(2)n′ |n′ = al,j+1;20: end for21: l∗ ← argmin{Temp(l)};22: if Temp(l∗) < s(2)max then23: Exchanging ai,j and al∗,j+1; // Confirm24: else25: flag = 0;26: end if27: end while

The GMSD algorithm is conducted in two phases.The first phase is initial scheduling. The resource blocks{Rn} are sequentially distributed to a one-dimensionalsequence, i.e., B1×LT in Algorithm 1. Then, reshape B toA.The second phase is optimal scheduling in a greedy

manner. It tries to decrease maximal second moment ofscheduling intervals iteratively. In each iteration, the sen-sor n with maximal second moment s(2)max is first selected.Then, find the largest scheduling interval of n and itsleft schedule ai,j. Try to exchange ai,j with its neighbor-ing schedules {al,j+1|l = 1, . . . , L} and record the minimalsecond moment from the sensors {n′|n′ = al,j+1, l =1, . . . , L}. If this minimal second moment obtained bysensor n′ = al∗,j+1 is smaller than s(2)max, confirm thescheduling exchange between n and n′; otherwise, cancelthe exchange and end the scheduling.

4.2 Separate design principle for priority-aware packetcollection

Based on the multichannel superframe design in previoussubsection, the SDP is obtained in this subsection.Figure 1 gives the illustration of SDP by an instance

with three priority classes and two wireless channels. Thenotation “nc” in the first sub-figure represents that the cor-responding slot on the specific channel is scheduled tothe sensor n from priority class c. As shown in the figure,SDP separately schedules the transmission for each classof sensors with the multichannel superframe, e.g., threesuperframes indicated by different color depths for threeclasses of sensors. With TDMA scheduling, the collisionsamong the sensors in the same priority class are avoided.However, as each resource block is scheduled to sensorsof different priority, it may lead to collisions when sen-sors of different priorities transmit packet simultaneously.As a solution, SDP introduces a priority-aware transmis-sion coordination mechanism by introducing the priorityindicator, which coordinates different sensors to transmitpacket in a given priority order. The details of SDP arepresented by the following two phases.1. Separate scheduling: Based on the multichannel

superframe presented in Section 4.1.1, we devise thesuperframe for each class of sensors separately. For thesuperframe of period T slots on L channels, the number ofresource blocks allocated to each sensor in priority classc is

Rn,c = λn,cUc

LT , n = 1, . . . ,Nc, c = 1, . . . ,C, (9)

whereUc = ∑Ncn=1 λn,c is the total packet rate of sensors in

priority class c. Note that there are two special cases due tothat each sensor work on a single channel. The first case isthat the number of sensors is smaller than the number of

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 6 of 14

Fig. 1 An instance of SDP-based multichannel scheduling for priority-aware packet collection: three priority classes and two channels

channels. In this case, each sensor is assigned to one chan-nel. The other case is that one sensor may be allocatedmore than T slots according to its proportional coeffi-cient. In this case, the sensor should be allocated T slotsand the remaining slots will be shared by others. Here, wefocus on the analysis of the general cases. The two specialcases can be analyzed in the similar way.For realization, we use the floor and ceil functions

to make sure that Rn,c is an integer. Then, the multi-channel superframes Ac for each priority class of sen-sors can be devised by GMSD algorithm separately andindependently.2. Priority-aware transmission coordination mechanism:

Since the transmission scheduling of each class of sensorsis designed by fully utilizing the resource blocks in thesuperframe, a priority-aware transmission coordinationmechanism is required to resolve the collisions among thesensors with different priorities. As shown in Fig. 1, weintroduce a priority indicator by setting C − 1 sub-slots

at the beginning of each time slot. The time length of asub-slot can be set to the backoff period defined in IEEE802.15.4 protocol. For the priority class c = 1, its sen-sors have the highest transmission priority. Hence, if onesensor of class 1 has data packet to transmit, it first trans-mits the carrier signal during the C − 1 sub-slots as thepriority indicator and then transmit the data packet dur-ing the remaining part of the slot. For any priority classc > 1, its sensors have to transmit data packet opportunis-tically. If one sensor from priority class c has data packetto transmit, it has to sense the channel for only one sub-slot, i.e., the sub-slot c − 1. If the channel is sensed idleduring this sub-slot, it transmits the carrier signal duringthe remaining sub-slots c, . . . ,C−1 and then transmits thedata packet; otherwise, it postpones the channel access tryto the next scheduled slot. Note that more complicatedpreamble signal can be used instead of the carrier signal,thus to embed the SDP into some existing communicationprotocols.

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 7 of 14

5 Performance analysis5.1 Mean transmission delay analysis based on SDPConsider the packet transmission process of each sensoras a queueing process. The transmission delay of each sen-sor is not only affected by the first and second momentsof scheduling intervals but also affected by the transmis-sion priority coordination in SDP. Note that if the packetarrival rates satisfy

∑Cc=1

∑Ncn=1 λn,c < L, the queueing

processes of all sensors are stable with the proportionalresource allocation by (9); thus, the packet transmission inthe network is stable. In the steady state, the probability ofthe scheduled resource block used by sensor n from prior-ity class c is the product of this sensor’s packet rate and itsmean scheduling interval, i.e.,

ρn,c = λn,csn,c = λn,cTRn,c

= UcL

= ρc, (10)

where sn,c is the mean scheduling interval of sensor n frompriority class c and ρc is the ratio of used network capac-ity by class c. According to the queueing theory, ρc is alsocalled utilization. Equation (10) implies that each resourceblock that is used by one sensor from priority class c isequivalent to the utilization of the priority class c, i.e., ρc.Recall the priority-aware transmission coordination

mechanism. For the highest priority c = 1, the sensorshave the absolute access right to their scheduled resourceblocks. That means if one sensor from priority class 1 haspackets to transmit during the scheduled slot, it can accessto the scheduled channel during the scheduled slot withprobability 1, denoted by p1 = 1 here. However, the occu-pation ratio of the resource by this sensor from priorityclass 1 is ρ1, i.e., the utilization of priority class 1. Forthe sensors from priority class 2, they have lower accessright than the sensors from priority class 1, which meansthat the sensors from priority class 2 can access to theirscheduled resource block with probability p2 = 1 − ρ1. Ingeneral, for sensors from priority class c, the occupationratio of their scheduled resource blocks by the higher pri-ority sensors is

∑c−1i=1 ρi; thus, the probability that they can

access to the scheduled resource blocks is

pc = 1 −c−1∑i=1

ρi, c = 1, . . . ,C. (11)

With the priority-aware transmission coordinationmechanism, the mean transmission interval of the sensorsfrom priority class c is equivalent to

s′n,c = sn,cpc

, (12)

and the lower bound of the second moment of transmis-sion interval is

s′n,c(2) =(sn,cpc

)2. (13)

As a result, the lower bound of mean waiting delay ofpacket transmission for sensor n from priority class c canbe obtained as

Wn,c= λn,cs′n,c(2)

2(1 − λn,cs′n,c

) = ρc2

2λn,c(1 − ∑c−1

i=1 ρi) (

1 − ∑ci=1 ρi

) .(14)

Since the successful transmission probability of a sensorfrom priority class c is pc, the mean time slots from theslot for this packet to be transmitted to the slot at which itis successfully transmitted can be given by

�n,c=pc ·1+∞∑k=1

(1−pc)kpc(1+ksn,c)= ρc∑c−1

i=1 ρi

λn,c(1−∑c−1

i=1 ρi) +1.

(15)

Hence, the lower bound of the mean packet transmis-sion delay of sensor n from priority class c is

Dn,c = ρc2

2λn,c(1 − ∑c−1

i=1 ρi) (

1 − ∑ci=1 ρi

)+ ρc

∑c−1i=1 ρi

λn,c(1 − ∑c−1

i=1 ρi) + 1.

(16)

5.2 Comparison with non-overlap schedulingFor general TDMA-based superframe design, in orderto avoid the collision, the resource block allocation isexclusive, such as GTS allocation in IEEE 802.15.4e andsuperframe design in WirelessHART. We call this type ofscheduling as non-overlap scheduling. In this subsection,we analyze the transmission delay performance of non-overlap scheduling and compare it to delay performanceof the SDP-based scheduling.For non-overlap scheduling, each resource block of the

superframe is scheduled to one sensor exclusively. In otherwords, the resource blocks of the superframe cannot befully scheduled to one priority class. Denote {αc|c =1, · · · ,C} as the weight coefficients of resource allocationto each priority class of sensors, where

∑Cc=1 αc = 1.

With proportional resource allocation in (1), the numberof resource blocks allocated to each sensor in priority classc is

Rn,c = λn,cUc

αcLT , n = 1, . . . ,Nc, c = 1, . . . ,C. (17)

where the tilde symbol is used to differ from the met-ric notations of SDP. Accordingly, the mean transmissioninterval of sensor n from priority class c is

sn,c = ρcλn,cαc

, (18)

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 8 of 14

and its corresponding lower bound of the second momentis

s (2)n,c = ( sn,c)2. (19)

As a result, the lower bound of the mean waiting delayof sensor n from priority class c can be obtained as

Wn,c = ρc2

2λn,cαc(αc − ρc), (20)

and the lower bound of the transmission delay is

Dn,c = ρc2

2λn,cαc(αc − ρc)+ 1. (21)

Based on the above analysis, we have the followingtheorem to demonstrate the advantage of SDP.

Theorem 1. For priority-ware packet collection, SDP-based scheduling achieves lower mean waiting delay foreach priority class than non-overlap scheduling.

Proof. Considering the stability of packet transmissionin non-overlap scheduling, it is easy to drive that ρc <

αc <

(1 − ∑

i∈[1,C],i�=cρi

). From (20), a larger αc results in

a smaller mean waiting delay. However, for ∀c ∈ [1,C], wehave

1 −∑

i∈[1,C], i�=cρi ≤ 1 −

c−1∑i=1

ρi, (22)

where the equality holds only when c = C, i.e., the lowest

priority class. As αc <

(1 − ∑

i∈[1,C],i�=cρi

), we can con-

clude that Wn,c < Wn,c for ∀c ∈[1,C]. Thus, Theorem 1holds.

Remark 1. Based on non-overlay scheduling, only thesensor in the lowest priority C can possibly acquire thesame waiting delay as the SDP-based scheduling does.However, in such case, the other priority classes are onlyallocated with resource at a portion of αc = ρc, c =1, · · · ,C − 1. That means the waiting delay of these classesof sensors will approach infinity. Although we theoreticallyhave Dn,c < Dn,c for certain class c given large αc, it willlead to rapid increase of the transmission delay of otherclasses. More deepgoing evaluations will be presented byperformance evaluation in the following section.

6 Simulation study and experimentsIn this section, we first evaluate themean packet transmis-sion delay of sensors for SDP-basedmultichannel schedul-ing. The comparison of simulated results and theoretical

lower bounds is presented. Then, we provide the compar-isons between SDP-based and IEEE 802.15.4e-based mul-tichannel scheduling. To show the feasibility of the SDP,the experiments on the SDR platform are also conducted.The network parameters for simulations are set as

shown in Table 1. The packet rates of sensors are ran-domly initialized but keep the utilization as shown in thetable. As IEEE 802.15.4e support multi-superframe, here,we set the period of multi-superframe as 16 × 2 slotwhereby the single superframe length is 16 slots.

6.1 Mean packet transmission delay evaluationThis part shows the simulation results of mean packettransmission delay of SDP-based multichannel schedul-ing and its corresponding theoretical lower bound, whichare marked by “SDP” and “SDP Theo,” respectively. Todemonstrate the influence of second moment of schedul-ing interval, we also conduct the simulation based onsequential scheduling (marked by “Seq”) with the mul-tichannel superframe obtained by initial scheduling inAlgorithm 1 (i.e., A at line 10).The simulation results are presented in Fig. 2. It

is shown that with simple sequential scheduling, themean packet transmission delay of each priority classis larger than that of the SDP-based scheduling. Usingthe GMSD, the mean packet transmission delay is effec-tively decreased and approaches to the theoretical lowerbound. Because some sensors, e.g., sensor 3 and sen-sor 4 in priority class 1, are allocated exact one channel,respectively, they have the same mean packet transmis-sion delay with different scheduling approaches. Further,we show the average delay of all sensors from same pri-ority class in Fig 3. As expected, the average transmissiondelay of each priority class increases with its priorityorder. However, compared to sequential scheduling, theincrement of SDP-based scheduling is relatively small.We also note that based on SDP, an increment of sen-sors brings in a small increment of the delay gap betweenthe simulated result and the theoretical result. That isbecause for a large number of sensors, it is non-trivial toapproach to the lower bound of mean packet transmissiondelay.

Table 1 Network parameters for simulation

Number of channels: L = 16

Period of superframe: T = 16 × 2

Number of priority classes: C = 4

Number of sensors in each priority class: N1 = 20, N2 = 30,

N3 = 40, N4 = 50

Utilization of each priority class: ρ1 = 0.1, ρ2 = 0.1,

ρ3 = 0.2, ρ4 = 0.2

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 9 of 14

2 4 6 8 10 12 14 16 18 200

1

2

3

4

5

Sensor from priority class 1

Mea

n de

lay

(slo

t)

5 10 15 20 25 300

5

10

Sensor from priority class 2

Mea

n de

lay

(slo

t)

5 10 15 20 25 30 35 400

5

10

15

Sensor from priority class 3

Mea

n de

lay

(slo

t)

5 10 15 20 25 30 35 40 45 500

5

10

15

20

25

Sensor from priority class 4

Mea

n de

lay

(slo

t)

Seq SDP TheoSDP

Seq SDP TheoSDP

Seq SDP TheoSDP

Seq SDP TheoSDP

Fig. 2Mean packet transmission delay of each sensor in four priority classes: SDP vs. sequential scheduling

1 1.5 2 2.5 3 3.5 40

2

4

6

8

10

12

14

Priority index

Ave

rage

del

ay (

slot

)

SeqSDPSDP Theo

Fig. 3 Average packet transmission delay of each priority class of sensors

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 10 of 14

6.2 SDP vs. non-overlap schedulingThis part shows the comparison between the SDP-basedscheduling and the non-overlap scheduling (marked by“NS”). As illustrated in Section 5.2, the transmission delayof sensors from different priority classes is affected bythe weighted coefficients {αc}. We first compare SDP toNS with moderate weighted coefficients α1 = 0.2, α2 =0.2, α3 = 0.3, and α4 = 0.3. The results shown inFig. 4 indicate that SDP achieves much lower mean packettransmission delay of each priority class of sensors thanthe NS does. The average delays of each class achieved bySDP are only 16, 18, 23, and 36 % of the average delaysachieved by NS, respectively.From (16) and (21), the mean packet transmission

delay of low-priority class based on NS can theoreti-cally be lower than the one based on SDP by adjust-ing the weighted coefficients. In the following, we checkthe case that the weighted coefficient α4 varying in itsrange ρ4 ∼

(1 − ∑3

i=1 ρi), i.e., 0.2 ∼ 0.6. The other

weighted coefficients are set as αi = (1−α4)ρi/(∑3

i=1 ρi)

for i = 1, 2, 3. The results are presented in Fig. 5.It is shown that the average delays of priority classes1, 2, and 3 based on SDP are lower than the onesbased on NS. As for priority class 4, the average delaybased on NS can hardly be lower than that based onSDP. Particularly when α4 ≥ 0.55, the average delayof priority class 4 based on NS is lower; however, theaverage delays of other high priority classes are dramati-cally increased (> 102 slots) and which is not acceptable.From simulation results, it can be concluded that theproposed SDP completely outperforms the non-overlapscheduling.

6.3 SDP vs. IEEE 802.15.4eFor comparison, we also evaluate the transmission delaywith multichannel scheduling method based on IEEE802.15.4e. As introduced in Section 2, in IEEE 802.15.4e,high-priority sensors are scheduled to transmit in CFP

2 4 6 8 10 12 14 16 18 200

5

10

15

Sensor from priority class 1

Mea

n de

lay

(slo

t) SDPNS

5 10 15 20 25 300

10

20

30

Sensor from priority class 2

Mea

n de

lay

(slo

t) SDPNS

5 10 15 20 25 30 35 400

10

20

30

Sensor from priority class 3

Mea

n de

lay

(slo

t) SDPNS

5 10 15 20 25 30 35 40 45 500

10

20

30

40

Sensor from priority class 4

Mea

n de

lay

(slo

t) SDPNS

Fig. 4Mean packet transmission delay of each sensor in four priority classes: SDP vs. non-overlap scheduling

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 11 of 14

0.25 0.3 0.35 0.4 0.45 0.5 0.5510

0

101

102

103

104

α4

Ave

rage

del

ay (

slot

)

NS: Priority 1NS: Priority 2NS: Priority 3NS: Priority 4SDP: Priority 1SDP: Priority 2SDP: Priority 3SDP: Priority 4

Fig. 5 Average packet transmission delay of each priority class: α4 varying from 0.25 to 0.55 in NS

10 20 30 40 50 60 70 80 90

101

102

103

104

105

106

Low priority sensor

Mea

n de

lay

(slo

t)

15.4e: CAP=1215.4e: CAP=11

SDP15.4e: CAP=815.4e: CAP=7

5 10 15 20 25 30 35 40 45 50

101

102

103

104

105

High priority sensor

Mea

n de

lay

(slo

t)

15.4e: CFP=415.4e: CFP=5

SDP15.4e: CFP=815.4e: CFP=9

(a) Mean packet transmission delays of high priority sensors

(b) Mean packet transmission delays of low priority sensorsFig. 6 Comparison results: mean packet transmission delay. aMean packet transmission delays of high-priority sensors. bMean packet transmissiondelays of low-priority sensors

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 12 of 14

4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 910

0

101

102

103

104

105

Length of CFP

Ave

rage

del

ay o

f sen

sors

(sl

ot)

15.4e: High priority15.4e: Low prioritySDP: High prioritySDP: Low priority

Fig. 7 Comparison results: average delay of all sensors with same priority

with TDMA manner and low-priority sensors are com-mitted to transmit in CAP with CSMA/CA manner. TheparametersNB, aMinBE, aMaxBE, andCW for CSMA areset to default values in IEEE 802.15.4. Here, we group thepriority classes 1 and 2 into one class, named high-priorityclass and group priority classes 3 and 4 into one class,named low-priority class. We partition the low-prioritysensors into 16 groups and assign each group with onechannel, thus to avoid the over crowded on some chan-nels. The CFP and CAP are tunable as used in [14]. Toguarantee the stability of packet transmission, under net-work settings in Table 1, we have 4 ≤ CFP ≤ 9 and7 ≤ CAP ≤ 12.Figure 6 compares the mean packet transmission delay

between SDP and IEEE 802.15.4e. From Fig. 6a, it can beseen that SDP, with full use of the multichannel super-frame, achieves lower mean packet transmission delayfor high-priority sensors than IEEE 802.15.4e. Since thesuperframe of IEEE 802.15.4e has to reserve CAP for low-priority sensors, the CFP is shorter compared to the oneof SDP. From the figure, when the CFP is small, e.g., 4slots, the mean packet transmission delay becomes verylarge. The comparison results of the packet transmissionof low priority sensors are in Fig. 6b. It is shown thatCSMA-based packet transmission results in large packettransmission delay even if the CAP is set relatively long.When the CAP is set small, the packet transmission isalmost blocked. The simulation results demonstrate thegreat advantage of SDP over the CSMA mechanism.The evaluation of the average delay of the sensors from

the same priority class is shown in Fig. 7. From the figure,it clearly shows that both the average delays of high-and low-priority sensors achieved by SDP are lower thanthe ones achieved by IEEE 802.15.4e. In addition, thepacket transmission delay performance of IEEE 802.15.4e

is even worse than non-overlap TDMA scheduling, forboth low- and high-priority classes of sensors. In sum-mary, for priority-aware packet collection in large-scaleindustrial WSNs, TDMA-based scheduling is better thanCSMA-based channel access. The proposed SDP providesthe lowest mean packet transmission delay.

6.4 Experiments on USRP platformTo demonstrate the feasibility of the proposed SDPscheme, preliminary experiments have been conducted onthe SDR platform implemented by USRPs (universal soft-ware radio peripherals). As shown in Fig. 8, five USRPs(version: NI USRP2921) are used during the experiments,where one is employed as the AP and the other four areused as the sensor nodes. Due to the limited devices, weconsider the scenario where two high-priority sensors andtwo low-priority sensors try to transmit the packets toone AP on one channel. The packet rates are set to 0.08and 0.12 packets per slot for two high-priority sensorsand 0.10 and 0.20 packets per slot for two low-priority

Fig. 8 SDR platform implemented by USRPs

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 13 of 14

1 2 3 40

5

10

15

20

25

30

Sensor index

Mea

n de

lay

(slo

t)

SDP TheoSDP ExpeNS Expe: α

1=0.3

NS Expe: α1=0.4

NS Expe: α1=0.5

NS Expe: α1=0.6

Fig. 9 Experiments results: mean packet transmission delay

sensors, respectively. As defined in IEEE 802.15.4, theOQPSK modulation/demodulation is used, the data rateis set to 250 Kbps and the bandwidth is 2 MHz. To avoidthe interference fromWiFi signal, the central frequency ofthe channel is set to 2425 MHz. The transmitting powerof the URSP is −10 dBm. Constrained to the processingspeed of the USRP, the time slot is set to 100 ms.On this SDR platform, the experiments for SDP- and

NS-based scheduling are conducted. Figure 9 presents theexperimental results after running 10,000 time slots foreach experiment.First, it is shown that the experimental results of the

mean packet transmission delay for each sensor based onSDP (denoted by “SDP Expe” in the figure) is close tothe theoretical lower bound. During the experiment, tworeasons make them slightly larger than the lower bound:(1) the scheduling based on superframe design is notstrictly periodic, which will increase the mean delay asproved by (5) in Section 4.1.2 and (2) the packet loss willoccur caused by the interference, and the retransmissionincreases the transmission delay.The comparison with the non-overlap scheduling

(denoted by “NS Expe” in the figure) is also presented inFig. 9. Let the weighted coefficients α1 for high-prioritysensors varies from 0.3 to 0.6, and α2 for low priority sen-sors is set to α2 = 1 − α1. Not surprisingly, the resultsdemonstrate that the SDP outperforms the non-overlapscheduling. It is shown when α1 = 0.3, more resource isallocated to the lower priority sensors. As a result, sensor3 gains a lower delay based on NS than that based on SDP;however, in this case, the delay of high-priority sensors areeven higher than the delay of low-priority sensors whichis not reasonable in application. When more resource isallocated to high-priority sensors, i.e., α1 > 0.3, it isshown that SDP achieves the lower mean transmission

delay than non-overlap scheduling for all of the foursensors.

7 ConclusionsIn this paper, a separate design principle-based mul-tichannel scheduling scheme is proposed for priority-aware packet collection in ICPS. The greedy multichannelscheduling determination algorithm is developed to opti-mize the design of the multichannel superframe. Apriority-aware packet transmission coordination mech-anism is devised to solve the channel access collision.Simulation results are provided to demonstrate thatthe SDP-based multichannel scheduling achieves lowermean transmission delay of each priority class than thatachieved by non-overlap scheduling and IEEE 802.15.4e-based scheduling. The feasibility of the proposed SDPis demonstrated through some preliminary experimentalresults. In future work, we will consider the extension ofSDP to a large-scale network topology for ICPS, such ascluter-tree topology.

Competing interestsThe authors declare that they have no competing interests.

AcknowledgementsThis work was supported in part by NSF of China under the grants 61221003,U1405251, 61290322, 61371085, 61431008, and 61273181, in part by NationalHigh Technology Research and Development Program of China (863 Program)under 2015AA01A702, in part by Ministry of Education of China underNCET-13-0358, and in part by Science and Technology Commission ofShanghai Municipality (STCSM), China under 13QA1401900.

Author details1Department of Automation, Shanghai Jiao Tong University, 800, DongchuanRd., Minhang District, 200240 Shanghai, People’s Republic of China. 2KeyLaboratory of Systems Control and Information Processing, Ministry ofEducation of China, 800, Dongchuan Rd., Minhang District, 200240 Shanghai,People’s Republic of China. 3School of Information Security Engineering,Shanghai Jiao Tong University, 800, Dongchuan Rd., Minhang District, 200240Shanghai, People’s Republic of China.

Lin et al. EURASIP Journal onWireless Communications and Networking (2016) 2016:71 Page 14 of 14

Received: 20 November 2015 Accepted: 24 February 2016

References1. F Xia, A Vinel, R Gao, L Wang, T Qiu, Evaluating IEEE 802.15.4 for

cyber-physical systems. EURASIP J. Wireless Commun. Netw.2011(596397), 1–14 (2011)

2. K-D Kim, PR Kumar, Cyber physical systems: a perspective at thecentennial. Proc. IEEE. 100(Special Centennial Issue), 1287–1308 (2012)

3. J Lee, B Bagheri, H-A Kao, A cyber-physical systems architecture forindustry 4.0-based manufacturing systems. Manuf. Lett. 3, 18–23 (2015)

4. PC Evans, M Annunziata, Industrial Internet: Pushing the Boundaries ofMinds andMachines. (General Electric Co., Schenectady, New York, 2012)

5. A Willig, E Uhlemann, Deadline-aware scheduling of cooperative relayersin TDMA-based wireless industrial networks. Wireless Netw. 20(1), 73–88(2014)

6. IEEE 802.15.4, Part 15.4: wireless medium access control (MAC) andphysical layer (PHY) specifications for low-rate wireless personal areanetworks (LR-WPANs). IEEE Standard for information technology (2006)

7. C Wang, T Jiang, Q Zhang, ZigBee® Network Protocols and applications.(CRC Press, Boca Raton, Florida, 2014)

8. IEC Std. 62,591, Industrial communication networks—wirelesscommunication network and communication profiles—WirelessHART.Int. Electrotech. Commission (2010)

9. ISA100.11a, Wireless systems for industrial automation: process controland related applications. ISA100 Standards Committee (2009)

10. IEC Std. 62,061, Industrial communication networks—fieldbusspecifications–WIA-PA communication networks and communicationprofiles. Int. Electrotech. Commission (2011)

11. G Anastasi, M Conti, M Di Francesco, A comprehensive analysis of theMAC unreliability problem in IEEE 802.15. 4 wireless sensor networks. IEEETrans. Ind. Inform. 7(1), 52–65 (2011)

12. IEEE 802.15.4e, IEEE standard for local and metropolitan areanetworks—part. 15.4: low-rate wireless personal area networks(LR-WPANs)—amendament1: MAC sublayer. IEEE Standard forinformation technology (2012)

13. F Lin, C Chen, C Hua, X Guan, in Proc. 10th International Conference onWireless Algorithms, Systems and Applications (WASA’15). SDP: separatedesign principle for multichannel scheduling in priority-aware packetcollection (Springer, Qufu, Shandong, China, 2015), pp. 356–365

14. MHS Gilani, I Sarrafi, M Abbaspour, An adaptive CSMA/TDMA hybrid MACfor energy and throughput improvement of wireless sensor networks. AdHoc Netw. 11(4), 1297–1304 (2013)

15. W Shen, T Zhang, F Barac, M Gidlund, PriorityMAC: a priority-enhancedMAC protocol for critical traffic in industrial wireless sensor and actuatornetworks. IEEE Trans. Ind. Inform. 10(1), 824–835 (2014)

16. A Zhou, M Liu, Z Li, E Dutkiewicz, Cross-layer design for proportional delaydifferentiation and network utility maximization in multi-hop wirelessnetworks. IEEE Trans. Wireless Commun. 11(4), 1446–1455 (2012)

17. U Bodin, K Wolosz, Proportional throughput differentiation with cognitiveload-control on WSN channels. EURASIP J. Wireless Commun. Netw.2015(1), 1–14 (2015)

18. J Liu, X Jiang, H Nishiyama, N Kato, X Shen, in Proc. 2012 IEEEWirelessCommunications and Networking Conference (WCNC’12). End-to-end delayin mobile ad hoc networks with generalized transmission range andlimited packet redundancy (IEEE, Shanghai, China, 2012), pp. 1731–1736

19. J Liu, X Jiang, H Nishiyama, N Kato, Generalized two-hop relay for flexibledelay control in MANETs. IEEE/ACM Trans. Netw. 20(6), 1950–1963 (2012)

20. J Liu, X Jiang, H Nishiyama, N Kato, On the delivery probability of two-hoprelay MANETs with erasure coding. IEEE Trans. Commun. 61(4),1314–1326 (2013)

21. E Toscano, L Lo Bello, Multichannel superframe scheduling for IEEE802.15. 4 industrial wireless sensor networks. IEEE Trans. Ind. Inform. 8(2),337–350 (2012)

22. A Saifullah, Y Xu, C Lu, Y Chen, End-to-end communication delay analysisin industrial wireless networks. IEEE Trans. Comput. 64(5), 1361–1374(2015)

23. A Azarfar, J-F Frigon, Sansò, Priority queueing models for cognitive radionetworks with traffic differentiation. EURASIP J. Wireless Commun. Netw.2014(1), 1–21 (2013)

24. N Zhang, H Liang, N Cheng, Y Tang, JW Mark, XS Shen, Dynamic spectrumaccess in multi-channel cognitive radio networks. IEEE J. Selected AreasCommun. 32(11), 2053–2064 (2014)

25. SM Ross, Introduction to Probability Models. (Academic Press, Waltham,Massachusetts, 2014)

Submit your manuscript to a journal and benefi t from:

7 Convenient online submission

7 Rigorous peer review

7 Immediate publication on acceptance

7 Open access: articles freely available online

7 High visibility within the fi eld

7 Retaining the copyright to your article

Submit your next manuscript at 7 springeropen.com